Warsztaty Badawcze 2021L - Praca Domowa nr 1

Autor: Bartosz Sawicki

Ten notatnik powstał w oparciu o https://www.kaggle.com/alpertml/credit-card-customers-eda-ml-97-5-accuracy/notebook#Exploring-the-Data

import numpy as np
import pandas as pd

import matplotlib.pyplot as plt
from matplotlib.ticker import MaxNLocator
import seaborn as sns
plt.style.use('ggplot')

from scipy import stats
from scipy.stats import norm


from sklearn.preprocessing import LabelEncoder 
from sklearn.preprocessing import OneHotEncoder
from sklearn.compose import ColumnTransformer 

import math
import warnings
warnings.filterwarnings('ignore')

input_df = pd.read_csv('BankChurners.csv')
display(input_df.shape)
display(input_df.sample(5))

input_df.drop('CLIENTNUM', axis=1, inplace=True)

input_df.drop(['Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_1'], 
          inplace=True, axis=1)

input_df.drop(['Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_2'], 
          inplace=True, axis=1)

display(input_df.shape)

(10127, 23)

	CLIENTNUM	Attrition_Flag	Customer_Age	Gender	Dependent_count	Education_Level	Marital_Status	Income_Category	Card_Category	Months_on_book	...	Credit_Limit	Total_Revolving_Bal	Avg_Open_To_Buy	Total_Amt_Chng_Q4_Q1	Total_Trans_Amt	Total_Trans_Ct	Total_Ct_Chng_Q4_Q1	Avg_Utilization_Ratio	Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_1	Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_2
5924	718469658	Existing Customer	37	F	2	Uneducated	Married	Unknown	Blue	22	...	7927.0	996	6931.0	0.521	4094	85	0.771	0.126	0.000311	0.99969
3494	796791258	Existing Customer	50	M	1	Uneducated	Married	$120K +	Blue	44	...	3104.0	1791	1313.0	1.048	4006	76	0.810	0.577	0.000315	0.99968
950	788814783	Existing Customer	39	M	1	Uneducated	Single	$60K - $80K	Blue	22	...	9204.0	845	8359.0	0.673	1820	58	0.611	0.092	0.000292	0.99971
5477	711379083	Existing Customer	60	F	0	Doctorate	Single	Less than $40K	Blue	42	...	2758.0	1429	1329.0	0.718	5124	74	0.644	0.518	0.000376	0.99962
5262	709486158	Existing Customer	33	F	3	Graduate	Married	Less than $40K	Blue	20	...	1438.3	0	1438.3	0.896	4530	72	0.636	0.000	0.000096	0.99990

5 rows × 23 columns

(10127, 20)

Kodowanie zmiennych kategorycznych i binarnych

updated_df = pd.DataFrame()
cats = ['Attrition_Flag', 'Gender', 'Education_Level', 'Marital_Status', 'Income_Category', 'Card_Category']

numeric_columns = ['Customer_Age','Credit_Limit','Months_on_book','Avg_Utilization_Ratio','Avg_Open_To_Buy','Total_Trans_Amt','Dependent_count',
                  'Total_Relationship_Count','Months_Inactive_12_mon','Contacts_Count_12_mon','Total_Revolving_Bal',
                  'Total_Amt_Chng_Q4_Q1','Total_Trans_Ct','Total_Ct_Chng_Q4_Q1']

def tobinary():
    
    # full_df['Attrition_Flag'] = full_df.Attrition_Flag // same thing
    updated_df['Attrition'] = input_df.Attrition_Flag.map({'Existing Customer':1, 'Attrited Customer':0})
    
    updated_df['Gender'] = input_df.Gender.map({'M':1, 'F':0})

def stringtoint():
    missing_income = input_df['Income_Category'].replace({'Unknown': 1 , 'Less than $40K':0, '$40K - $60K':0, 
                                                      '$80K - $120K':0, '$60K - $80K':0, '$120K +':0})
    #missinng data will be replaced with mode:
    income_data    = input_df['Income_Category'].replace({'Unknown': 1 , 'Less than $40K':1, '$40K - $60K':2, 
                                                      '$80K - $120K':3, '$60K - $80K':4, '$120K +':5})
    
    
    
    missing_education = input_df['Education_Level'].replace({'Unknown': 1, 'High School':0, 'Graduate':0, 'Uneducated':0,
                                                         'College':0,'Post-Graduate':0,'Doctorate':0})
    #missinng data will be replaced with mode:
    education_data    = input_df['Education_Level'].replace({'Unknown': 2, 'High School':1, 'Graduate':2, 'Uneducated':3,
                                                         'College':4,'Post-Graduate':5,'Doctorate':6})
    
    updated_df['Missing_Income']    = missing_income
    updated_df['Income_Category']   = income_data
    updated_df['Missing_Education'] = missing_education
    updated_df['Education_Level']   = education_data

def encode():
    global updated_df
    card_dummies = pd.get_dummies(input_df['Card_Category'], prefix='Card')
    marital_dummies = pd.get_dummies(input_df['Marital_Status'], prefix='Marital')
    updated_df = pd.concat([updated_df, marital_dummies, card_dummies], axis=1)

def concat_with_numerics():
    global updated_df
    updated_df = pd.concat([updated_df, input_df.loc[:, numeric_columns]], axis=1)

tobinary()
stringtoint()
encode()
concat_with_numerics()

print('Data shapes """including target value"""')
print(f'Old shape : {input_df.shape}')
print(f'Updated shape : {updated_df.shape}')

Data shapes """including target value"""
Old shape : (10127, 20)
Updated shape : (10127, 28)

Modelowanie

import xgboost as xgb
from sklearn.model_selection import StratifiedKFold, cross_validate

def estimates(X_data, y_data, models, cv):
    
    train_acc_dict = dict()
    test_acc_dict = dict()
    time_dict = dict()
    
    for model in models:
        
        current_model_name = model.__class__.__name__
        
        cv_results = cross_validate(model, X_data, y_data, cv=cv,
                                    return_train_score=True, scoring='accuracy')
        
        train_acc_dict[current_model_name] = cv_results['train_score'].mean()
        test_acc_dict[current_model_name] = cv_results['test_score'].mean()
        time_dict[current_model_name] = cv_results['fit_time'].mean()
        
    return train_acc_dict, test_acc_dict, time_dict

m_xgb = xgb.XGBClassifier(n_estimators=2250,
                       max_depth=2, random_state=14)

cv = StratifiedKFold(11, shuffle=True, random_state=14)

models = [m_xgb]

X = updated_df.drop('Attrition', axis=1)
y = updated_df['Attrition']

print(X.shape)
print(y.shape)

train_acc_dict, test_acc_dict, time_dict = estimates(X, y, models, cv)

(10127, 27)
(10127,)
[14:15:19] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:15:25] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:15:40] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:15:46] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:15:52] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:15:57] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:04] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:10] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:16] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:33] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:39] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.

# Test accuracy
for key, value in test_acc_dict.items():
    print('{} - {:.1f}%'.format(key, value*100))

XGBClassifier - 97.1%

Sprawdźmy jak radzi sobie gdy usuniemy skorelowane chechy i dobierzemy inne hiperparametry.

X_small = X.drop(['Customer_Age', 'Total_Trans_Ct', 'Credit_Limit'], axis='columns')

deeper_xgb =  xgb.XGBClassifier(n_estimators=100,
                       max_depth=5, random_state=14)

cv = StratifiedKFold(11, shuffle=True, random_state=14)

xgb_clfs = [deeper_xgb]

train_acc_dict, test_acc_dict, time_dict = estimates(X_small, y, xgb_clfs, cv)
test_acc_dict

[14:16:46] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:46] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:47] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:48] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:48] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:49] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:50] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:50] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:51] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:51] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.
[14:16:52] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.

{'XGBClassifier': 0.9619847390488943}

Podział zbioru na treningowy i testowy + trenowanie modelu

from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(X_small, y, random_state=14)

deeper_xgb.fit(X_train, y_train)

[14:16:53] WARNING: ../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.

XGBClassifier(base_score=0.5, booster='gbtree', colsample_bylevel=1,
              colsample_bynode=1, colsample_bytree=1, gamma=0, gpu_id=-1,
              importance_type='gain', interaction_constraints='',
              learning_rate=0.300000012, max_delta_step=0, max_depth=5,
              min_child_weight=1, missing=nan, monotone_constraints='()',
              n_estimators=100, n_jobs=4, num_parallel_tree=1, random_state=14,
              reg_alpha=0, reg_lambda=1, scale_pos_weight=1, subsample=1,
              tree_method='exact', validate_parameters=1, verbosity=None)

Wyjaśnianie

!pip install dalex -U

Requirement already up-to-date: dalex in /home/sawcio/Studia/4sem/Warsztaty_badawcze/wb-env/lib/python3.8/site-packages (1.0.1)
Requirement already satisfied, skipping upgrade: plotly>=4.12.0 in /home/sawcio/Studia/4sem/Warsztaty_badawcze/wb-env/lib/python3.8/site-packages (from dalex) (4.14.3)
Requirement already satisfied, skipping upgrade: pandas>=1.1.2 in /home/sawcio/Studia/4sem/Warsztaty_badawcze/wb-env/lib/python3.8/site-packages (from dalex) (1.2.2)
Requirement already satisfied, skipping upgrade: setuptools in /home/sawcio/Studia/4sem/Warsztaty_badawcze/wb-env/lib/python3.8/site-packages (from dalex) (47.1.0)
Requirement already satisfied, skipping upgrade: tqdm>=4.48.2 in /home/sawcio/Studia/4sem/Warsztaty_badawcze/wb-env/lib/python3.8/site-packages (from dalex) (4.59.0)
Requirement already satisfied, skipping upgrade: numpy>=1.18.4 in /home/sawcio/Studia/4sem/Warsztaty_badawcze/wb-env/lib/python3.8/site-packages (from dalex) (1.20.1)
Requirement already satisfied, skipping upgrade: six in /home/sawcio/Studia/4sem/Warsztaty_badawcze/wb-env/lib/python3.8/site-packages (from plotly>=4.12.0->dalex) (1.15.0)
Requirement already satisfied, skipping upgrade: retrying>=1.3.3 in /home/sawcio/Studia/4sem/Warsztaty_badawcze/wb-env/lib/python3.8/site-packages (from plotly>=4.12.0->dalex) (1.3.3)
Requirement already satisfied, skipping upgrade: pytz>=2017.3 in /home/sawcio/Studia/4sem/Warsztaty_badawcze/wb-env/lib/python3.8/site-packages (from pandas>=1.1.2->dalex) (2021.1)
Requirement already satisfied, skipping upgrade: python-dateutil>=2.7.3 in /home/sawcio/Studia/4sem/Warsztaty_badawcze/wb-env/lib/python3.8/site-packages (from pandas>=1.1.2->dalex) (2.8.1)
WARNING: You are using pip version 20.1.1; however, version 21.0.1 is available.
You should consider upgrading via the '/home/sawcio/Studia/4sem/Warsztaty_badawcze/wb-env/bin/python -m pip install --upgrade pip' command.

import dalex as dx

explainer = dx.Explainer(deeper_xgb, X_train, y_train)

Preparation of a new explainer is initiated

  -> data              : 7595 rows 24 cols
  -> target variable   : Parameter 'y' was a pandas.Series. Converted to a numpy.ndarray.
  -> target variable   : 7595 values
  -> model_class       : xgboost.sklearn.XGBClassifier (default)
  -> label             : Not specified, model's class short name will be used. (default)
  -> predict function  : <function yhat_proba_default at 0x7f69b9f41af0> will be used (default)
  -> predict function  : Accepts pandas.DataFrame and numpy.ndarray.
  -> predicted values  : min = 9.37e-06, mean = 0.841, max = 1.0
  -> model type        : classification will be used (default)
  -> residual function : difference between y and yhat (default)
  -> residuals         : min = -0.779, mean = -5.26e-05, max = 0.7
  -> model_info        : package xgboost

A new explainer has been created!

display(explainer.predict(X_test)[4], y_test.iloc[4])

0.9999571

1

pp = explainer.predict_parts(X_test.iloc[4,:])

pp

	variable_name	variable_value	variable	cumulative	contribution	sign	position	label
0	intercept	1	intercept	0.840606	8.406057e-01	1.0	23	XGBClassifier
1	Total_Trans_Amt	14240.0	Total_Trans_Amt = 14240.0	0.986708	1.461025e-01	1.0	22	XGBClassifier
2	Total_Relationship_Count	2.0	Total_Relationship_Count = 2.0	0.987994	1.286149e-03	1.0	21	XGBClassifier
3	Total_Revolving_Bal	1465.0	Total_Revolving_Bal = 1465.0	0.990510	2.515256e-03	1.0	20	XGBClassifier
4	Months_Inactive_12_mon	1.0	Months_Inactive_12_mon = 1.0	0.996665	6.155789e-03	1.0	19	XGBClassifier
5	Total_Amt_Chng_Q4_Q1	0.63	Total_Amt_Chng_Q4_Q1 = 0.63	0.999335	2.669394e-03	1.0	18	XGBClassifier
6	Total_Ct_Chng_Q4_Q1	0.656	Total_Ct_Chng_Q4_Q1 = 0.656	0.999560	2.249479e-04	1.0	17	XGBClassifier
7	Avg_Utilization_Ratio	0.407	Avg_Utilization_Ratio = 0.407	0.999497	-6.252527e-05	-1.0	16	XGBClassifier
8	Contacts_Count_12_mon	2.0	Contacts_Count_12_mon = 2.0	0.999944	4.463196e-04	1.0	15	XGBClassifier
9	Months_on_book:Marital_Married	48.0:1.0	Months_on_book:Marital_Married = 48.0:1.0	0.999951	7.748604e-06	1.0	14	XGBClassifier
10	Dependent_count	3.0	Dependent_count = 3.0	0.999960	8.761883e-06	1.0	13	XGBClassifier
11	Gender	0.0	Gender = 0.0	0.999960	1.192093e-07	1.0	12	XGBClassifier
12	Missing_Education	1.0	Missing_Education = 1.0	0.999954	-6.496906e-06	-1.0	11	XGBClassifier
13	Avg_Open_To_Buy:Marital_Single	2138.0:0.0	Avg_Open_To_Buy:Marital_Single = 2138.0:0.0	0.999960	6.556511e-06	1.0	10	XGBClassifier
14	Education_Level	2.0	Education_Level = 2.0	0.999956	-4.291534e-06	-1.0	9	XGBClassifier
15	Income_Category	1.0	Income_Category = 1.0	0.999959	2.682209e-06	1.0	8	XGBClassifier
16	Marital_Divorced	0.0	Marital_Divorced = 0.0	0.999957	-1.609325e-06	-1.0	7	XGBClassifier
17	Card_Blue	1.0	Card_Blue = 1.0	0.999957	0.000000e+00	0.0	6	XGBClassifier
18	Card_Gold	0.0	Card_Gold = 0.0	0.999957	0.000000e+00	0.0	5	XGBClassifier
19	Card_Silver	0.0	Card_Silver = 0.0	0.999957	0.000000e+00	0.0	4	XGBClassifier
20	Missing_Income	0.0	Missing_Income = 0.0	0.999957	0.000000e+00	0.0	3	XGBClassifier
21	Marital_Unknown	0.0	Marital_Unknown = 0.0	0.999957	0.000000e+00	0.0	2	XGBClassifier
22	Card_Platinum	0.0	Card_Platinum = 0.0	0.999957	0.000000e+00	0.0	1	XGBClassifier
23			prediction	0.999957	9.999571e-01	1.0	0	XGBClassifier

pp.plot()

pp_shap = explainer.predict_parts(X_train.iloc[4,:], type='shap')

pp_shap

	variable	contribution	variable_name	variable_value	sign	label	B
0	Card_Silver = 0.0	-0.000042	Card_Silver	0.0	-1.0	XGBClassifier	1
1	Missing_Education = 0.0	0.000313	Missing_Education	0.0	1.0	XGBClassifier	1
2	Contacts_Count_12_mon = 2.0	0.007205	Contacts_Count_12_mon	2.0	1.0	XGBClassifier	1
3	Income_Category = 1.0	0.000376	Income_Category	1.0	1.0	XGBClassifier	1
4	Gender = 0.0	-0.003118	Gender	0.0	-1.0	XGBClassifier	1
...	...	...	...	...	...	...	...
19	Marital_Single = 1.0	-0.000071	Marital_Single	1.0	-1.0	XGBClassifier	0
20	Card_Blue = 1.0	0.000060	Card_Blue	1.0	1.0	XGBClassifier	0
21	Missing_Income = 0.0	-0.000015	Missing_Income	0.0	-1.0	XGBClassifier	0
22	Card_Silver = 0.0	-0.000015	Card_Silver	0.0	-1.0	XGBClassifier	0
23	Card_Platinum = 0.0	0.000000	Card_Platinum	0.0	0.0	XGBClassifier	0

624 rows × 7 columns

pp_shap.plot(max_vars=16)

• Widzimy, że największy wpływ na predykcję miała zmienna Total_Trans_Amt. Spójrzmy na jej gęstość z podziałem na zmienną celu: Widać, że ta zmienna dosyć dobrze różnicuje klientów, którzy odeszli od tych którzy zostali.

• Wyniki Break Down i Shapley Values różnią się od siebie, w szczególności:

  • Shapley podaje jako 2. najważniejszą wartość Total_Ct_Chng_Q4_Q1, a dla Break Down wkład tej zmiennej wynosi 0.
  • Według Break Down Total_Trans_Amt przyczyniła się o 0.146 i zdecydowanie ma największy wpływ na wynik, a Shapley Values rozdzielaja wkład na więcej zmiennych.
• Analiza Shap trwa dłużej niż Break Down

Szukanie obserwacji o różnych najważniejszych zmiennych

for index,observation in enumerate(X_test.itertuples()):
    pp = explainer.predict_parts(pd.DataFrame(observation).drop([0]).transpose(), type = 'break_down')
    _vars = [pp.result.loc[1,:].variable_name, pp.result.loc[2,:].variable_name]
    display(index)
    display(_vars)
    if index == 10:
        break

0

['Total_Amt_Chng_Q4_Q1', 'Months_Inactive_12_mon']

1

['Total_Trans_Amt', 'Total_Ct_Chng_Q4_Q1']

2

['Total_Revolving_Bal', 'Total_Amt_Chng_Q4_Q1']

3

['Total_Revolving_Bal', 'Total_Ct_Chng_Q4_Q1']

4

['Total_Trans_Amt', 'Total_Revolving_Bal']

5

['Total_Trans_Amt', 'Months_Inactive_12_mon']

6

['Total_Amt_Chng_Q4_Q1', 'Contacts_Count_12_mon']

7

['Total_Trans_Amt', 'Total_Revolving_Bal']

8

['Total_Amt_Chng_Q4_Q1', 'Avg_Open_To_Buy']

9

['Total_Amt_Chng_Q4_Q1', 'Months_on_book']

10

['Total_Amt_Chng_Q4_Q1', 'Contacts_Count_12_mon']

pp = explainer.predict_parts(X_test.iloc[5,:], type = 'break_down')
pp.plot()
y_test.iloc[5]

1

pp = explainer.predict_parts(X_test.iloc[6,:], type = 'break_down')
pp.plot()
y_test.iloc[6]

0

index	5	6
1. najważniejsza zmienna	Total_Trans_Amt	Total_Amt_Chng_Q4_Q1
2. najważniejsza zmienna	Months_Inactive_12_mon	Contacts_Count_12_mon
predykcja i etykieta	klient pozostał	klient odszedł

Te obserwacje nie tylko różnią się najważniejszymi zmiennymi, ale też predykcją i etykietą. Może to ma wpływ na wybór najważniejszych cech przez metodę Break Down. Spróbujmy wyjaśnić kilka obserwacji, które mają etykietę 0.

attrited = X_test[y_test==0].sample(5, random_state=997)
explainer.predict_parts(attrited.iloc[0,:], type = 'break_down').plot(max_vars=24)
explainer.predict_parts(attrited.iloc[1,:], type = 'break_down').plot(max_vars=24)
explainer.predict_parts(attrited.iloc[2,:], type = 'break_down').plot(max_vars=24)

Wśród próbki z obserwacji z etykietą 0 różne cechy występują jako najważniejsze. Kolejność cech nie zależy więc od etykiety i predykcji.

Obserwacje, które dla tych samych zmiennych mają inne efekty

pp = explainer.predict_parts(X_test.iloc[55,:], type = 'break_down')
pp.plot(max_vars=24)

pp = explainer.predict_parts(X_test.iloc[122,:], type = 'break_down')
pp.plot(max_vars=24)

Dla 55 i 122 obserwacji cecha Contacts_Count_12_mon jest równa 3, jednak w pierwszym przypadku zwiększa predykcję o 0.003, a w drugim zmniejsza o 0.006.

Wnioski ogólne

Cechy takie jak rodzaj karty czy stan cywilny wpływają nieznacznie lub zupełnie nie wpływają na predykcję modelu. Możnaby spróbować usunąć je i sprawdzić jaką skuteczność osiąga model.